Agentic automation
What is an n8n AI agent?
n8n documents what its AI Agent node does, how it differs from a fixed chain, what limits it has and where a person can approve a tool call. This page reports that in plain words.
An n8n AI agent is a workflow built around the AI Agent node. n8n says you connect a chat model and one or more tools, and the agent decides which tools to call to complete a task. A plain n8n workflow runs the steps you drew in the order you drew them. An agent chooses its own next step, which is useful for open tasks and is the reason it needs limits.
plain answer
An n8n AI agent is a workflow where a model chooses the tools
n8n's documentation for the AI Agent node says: "The AI Agent node lets you build an AI agent in n8n." It describes the setup as connecting "a chat model and one or more tools, and the agent decides which tools to call to complete a task."
The same page says the agent "can understand the capabilities of different tools and determine which tool to use depending on the task." The node needs at least one tool. n8n's page says: "You must connect at least one tool sub-node to an AI Agent node." For how this fits into a wider service, see our page on agentic AI automation services.
plain answer
An agent differs from a chain because it decides what to do next
n8n's own explainer puts it this way: "Where a chain follows a predetermined sequence of calls to different AI components," "an agent uses a language model to determine which actions to take."
n8n's comparison page describes the Basic LLM Chain node as one that "supports chatting with a connected LLM, but doesn't support memory or tools." So a chain suits a task whose steps you already know, and an agent suits a task where the steps depend on the input. That is our reading of n8n's comparison, not n8n's wording. Our guide to agentic AI versus traditional automation covers the same split outside n8n.
plain answer
One agent run, step by step
n8n describes the shape of a run like this: "When you execute a workflow containing an agent, the agent runs multiple times. For example, it may do an initial setup, followed by a run to call a tool, then another run to evaluate the tool response and respond to the user."
Here is an invented example that follows that shape. It is an illustration written by us, not a recorded run and not a test result. Suppose a support workflow starts from a chat message and the agent has two tools connected: one that looks up an order in a shipping system, and one that sends an email.
The message reads: "Has order 1042 shipped?" In the first run, the model receives the message, its system message and the list of connected tools. It decides that the lookup tool fits and chooses the order number as the input. In the second run, the lookup tool has returned a status, and the model reads it. It does not need the email tool, so it does not call it. In the final run, the model writes the reply to the customer from the status it was given.
n8n's Tools Agent page documents an option for seeing this afterwards: Return Intermediate Steps, which it describes as selecting "whether to include intermediate steps the agent took in the final output (turned on) or not (turned off)." n8n adds that it "could be useful for further refining the agent's behavior based on the steps it took."
plain answer
What happens when no tool fits is not stated in the pages we read
The AI Agent node page and the Tools Agent page, both read on 18 November 2026, do not say what the node does when none of the connected tools fits the request. We did not test it on an n8n instance, so this page does not claim a behavior.
What n8n does document are the settings that shape the decision. The System Message option is described as a way to "guide the agent's decision-making." The Max Iterations option is described as the number of times the model should run "to try and generate a good answer from the user's prompt," and n8n says it "Defaults to 10."
Our reading, not n8n's: write into the system message what the agent should do when no tool applies, then test with a request that no connected tool can handle and read the intermediate steps to see what the agent actually did.
plain answer
The limits n8n documents for an agent
Several limits appear in n8n's own pages. An agent needs at least one tool (S1). Its model runs are set by the Max Iterations option, which defaults to 10 (S5). Memory is optional and attached as a separate node: n8n says to "Attach a memory sub-node so that users can have an ongoing conversation with multiple queries," and also says "Memory doesn't persist between sessions" (S5). Its Simple Memory option "stores a customizable length of chat history for the current session" (S3). We did not test how long any stored history is kept in a given setup, so this page makes no claim about that.
A separate source describes why the tool list matters. The OWASP Gen AI Security Project names Excessive Agency as a risk and says its root cause is "typically one or more of: excessive functionality; excessive permissions; excessive autonomy." Among its prevention steps it lists: "Limit the extensions that LLM agents are allowed to call to only the minimum necessary." In n8n terms, our reading is that each tool you connect is something the model may choose to call, so the tool list is part of the agent's limits. Our page on AI agent security risk covers how to review what an agent can reach.
limits
n8n lets you require a person to approve a tool call
n8n's documentation says: "You can require human approval before an AI Agent executes a specific tool." It says the review can apply "to all tools connected to an AI Agent node, or just selected individual tools," and that it is aimed at higher-risk tools "such as sending messages, modifying records, or deleting data."
On its page, n8n says the reviewer receives "the request showing which tool the AI wants to use and with what parameters." If approved, "the tool executes with the AI-specified input." If denied, "the action is canceled and the AI is informed of the rejection." The approval request can go through a channel such as Slack, Telegram or the n8n Chat interface, and n8n lists Microsoft Teams and email options as well.
n8n also tells builders to describe the review setup in the agent's system prompt, so the agent knows which tools need approval and how to respond when a call is denied. In the support example above, that would mean putting the email tool behind a review step and leaving the lookup tool without one. That choice is ours, not n8n's. Our article on approvals in n8n covers how to design an approval programme for automations in general; it does not cover agent tool calls.
Sources
Cited sources
- S1: AI Agent node (n8n Docs), read 2026-11-18
- S2: What agents do (n8n Docs), read 2026-11-18
- S3: How memory works (n8n Docs), read 2026-11-18
- S4: Human-in-the-loop for tools (n8n Docs), read 2026-11-18
- S5: Tools Agent (n8n Docs), read 2026-11-18
- S6: Agents vs chains (n8n Docs), read 2026-11-18
- S7: LLM06:2025 Excessive Agency (OWASP Gen AI Security Project), read 2026-11-18
Related resources
Continue exploring
Straight answers
Frequently asked questions about What is an n8n AI agent?
What is an n8n AI agent?
A workflow built around n8n's AI Agent node. n8n says you connect a chat model and one or more tools, and the agent decides which tools to call to complete a task. Source: S1, read 2026-11-17.
Does the n8n AI Agent node need a tool?
Yes. n8n's page says you must connect at least one tool sub-node to an AI Agent node. Source: S1, read 2026-11-17.
What does an n8n AI agent do when no tool fits the request?
The n8n pages we read do not say, and we did not test it. n8n documents a System Message option to guide the agent's decision-making and a Max Iterations option that defaults to 10. Source: S5, read 2026-11-17.
Can an n8n AI agent remember earlier messages?
n8n says to attach a memory sub-node for an ongoing conversation, and that memory does not persist between sessions. Its Simple Memory option stores chat history for the current session. Sources: S5 and S3, read 2026-11-17. We did not test retention.
Can a person approve what an n8n AI agent does?
n8n documents a human approval step that can be required before an AI Agent runs a specific tool: the workflow pauses and a reviewer sees the tool and its parameters. Source: S4, read 2026-11-17.
Is an n8n AI agent the same as a normal n8n workflow?
No. A chain follows a predetermined sequence, while n8n says an agent uses a language model to determine which actions to take. Source: S2, read 2026-11-17.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call